Accel-Sim 2.0 — Validated GPU Simulation with full Hopper support
Accel-Sim is an extensible, validated framework for cycle-level GPU simulation.
It traces real SASS execution from NVIDIA hardware with NVBit and replays it on a detailed performance model (GPGPU-Sim 4.x), together with the AccelWattch power model.
Accel-Sim 2.0 adds full NVIDIA Hopper (H100 / H200) support: the Tensor Memory Accelerator (TMA), asynchronous Warp Group MMA (WGMMA), mbarrier producer/consumer synchronization, threadblock clusters, a chiplet / uGPU partitioned memory subsystem, a rebuilt tracer (compressed traces + PyTorch per-layer hooks), and GPUVision for cycle-level hardware correlation.
See the full changelog and Table 1 in release.notes.md.
🚀 Simulate contemporary LLM inference and training, end to end
Accel-Sim 2.0 fully supports vLLM inference and PyTorch training — grab any model from
Hugging Face, trace inference or a full training step (forward + backward + optimizer) with
a few lines of Python, and run it on a cycle-level model of an H100. No hand-written
kernels, no unsupported-op workarounds: modern async Hopper kernels (FlashAttention-3,
cuBLAS/CUTLASS, NCCL) run out of the box.
And it's accurate.
Validated across 34,000+ kernel instances from 22 benchmark suites, Accel-Sim 2.0 achieves a 99% Pearson correlation and just 13.4% mean absolute cycle error against real NVIDIA H100 silicon.
→ Jump to Tracing LLMs with vLLM.
Simulated vs. real H100 across 34,000+ kernels — GPC cycles, warp instructions, and L1/L2 accesses & misses.
Each panel shows the Pearson correlation and mean absolute percentage error (MAPE) for that metric.
Roadmap: experimental Blackwell (B200 / RTX 5090) and multi-GPU / NVLink support are incoming shortly.
Contents
How to Cite
Cite Accel-Sim cumulatively.
If you use anything in this repo, cite the base papers;
add the power papers on top only if you use AccelWattch or GPUWattch.
The Accel-Sim 2.0 paper introduces full Hopper support and the 2.0 framework;
the ISCA 2020 paper introduces Accel-Sim;
the MICRO 2021 paper introduces AccelWattch (see the AccelWattch MICRO'21 Artifact Manual).
Ready-to-use metadata ships with the repo: CITATION.bib (every entry below, copy-pasteable) and CITATION.cff (machine-readable; drives GitHub's "Cite this repository" button in the sidebar).
Note that GitHub's button exports only the Accel-Sim 2.0 entry — use CITATION.bib for the full cumulative set.
1. Base — always cite these
Required for any use of Accel-Sim: the tracer, the trace-driven frontend, the tuner, the correlator, the Hopper model — anything.
Accel-Sim 2.0 (2026) ＋ Accel-Sim 1.0 (ISCA'20) ＋ GPGPU-Sim (ISPASS'09)
@misc{accelsim2_2026,
author = {Junrui Pan and Weili An and Cesar Avalos Baddouh and Christin David Bose and
Ni Kang and Aaron Barnes and Ahmad Alawneh and Fangjia Shen and Yechen Liu and
Anusuya Nallathambi and Atthin Chandrashekar and Timothy G. Rogers},
title = {Architecting the Next Generation of Asynchronous, Distributed {GPUs} for the {AI} Era},
year = {2026},
eprint = {2608.22602},
archivePrefix = {arXiv},
primaryClass = {cs.AR},
doi = {10.48550/arXiv.2608.22602},
url = {https://arxiv.org/abs/2608.22602}
}
@inproceedings{accelsim2020,
author = {Mahmoud Khairy and Zhesheng Shen and Tor M. Aamodt and Timothy G. Rogers},
title = {Accel-Sim: An Extensible Simulation Framework for Validated {GPU} Modeling},
booktitle = {ISCA},
year = {2020},
doi = {10.1109/ISCA45697.2020.00047}
}
@inproceedings{gpgpusim2009,
author = {Ali Bakhoda and George L. Yuan and Wilson W. L. Fung and Henry Wong and Tor M. Aamodt},
title = {Analyzing {CUDA} Workloads Using a Detailed {GPU} Simulator},
booktitle = {ISPASS},
year = {2009},
doi = {10.1109/ISPASS.2009.4919648}
}
2. Power modeling — add these if you use AccelWattch or GPUWattch
These are in addition to the base papers above, not instead of them.
＋ AccelWattch (MICRO'21) ＋ GPUWattch (ISCA'13)
This is the beginning of the README. Read the rest at the original. Read the full README ↗
